EgoSAT - Multi-Step Recall: leaderboard
EgoSAT (2026) tests streaming understanding of egocentric video: at each query time the model sees only the video so far and answers four-option questions built from Ego4D annotations. This board asks about past actions at one, two and three steps back (mean accuracy of the three probes); accuracy in percent, chance is 25 and higher is better.
Metric: Multiple-choice accuracy (%) on past actions at one, two and three steps back (mean accuracy of the three probes); EgoSAT: streaming questions on Ego4D egocentric video answered online from the clips seen so far (strict prefix truncation at each query time); four options per question (the answer, two temporally adjacent hard negatives and one absurd negative), chance 25; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro | 48.15 | #266 |
| 2 | Qwen 2.5 VL 72B Instruct | 43.46 | #846 |
| 3 | Qwen 2.5 VL 32B Instruct | 40.74 | #1053 |
| 4 | Claude Sonnet 4 | 39.64 | #400 |
| 5 | Qwen 2.5 VL 7B Instruct | 34.46 | #1833 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=egosat-multi-step-recall · How It Works · Data refreshed daily, snapshot 2026-11-13.